{"bundle_type":"pith_open_graph_bundle","bundle_version":"1.0","pith_number":"pith:2026:M7LQKF3NPAOBGIVK7FBGNA3JSQ","short_pith_number":"pith:M7LQKF3N","canonical_record":{"source":{"id":"2605.08879","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.RO","submitted_at":"2026-05-09T10:59:03Z","cross_cats_sorted":[],"title_canon_sha256":"c53e7388114353dc17753acb434f7799b4563abced5c0edd7c6d3f6742166908","abstract_canon_sha256":"c69278ea19f1bb98ffe8f9e7c40c0ec650dbcd746b7299449bd7e919e49c780e"},"schema_version":"1.0"},"canonical_sha256":"67d705176d781c1322aaf942668369942bd007edb13abc41477feb2956d4361c","source":{"kind":"arxiv","id":"2605.08879","version":2},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.08879","created_at":"2026-05-20T01:05:16Z"},{"alias_kind":"arxiv_version","alias_value":"2605.08879v2","created_at":"2026-05-20T01:05:16Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.08879","created_at":"2026-05-20T01:05:16Z"},{"alias_kind":"pith_short_12","alias_value":"M7LQKF3NPAOB","created_at":"2026-05-20T01:05:16Z"},{"alias_kind":"pith_short_16","alias_value":"M7LQKF3NPAOBGIVK","created_at":"2026-05-20T01:05:16Z"},{"alias_kind":"pith_short_8","alias_value":"M7LQKF3N","created_at":"2026-05-20T01:05:16Z"}],"events":[{"event_type":"record_created","subject_pith_number":"pith:2026:M7LQKF3NPAOBGIVK7FBGNA3JSQ","target":"record","payload":{"canonical_record":{"source":{"id":"2605.08879","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.RO","submitted_at":"2026-05-09T10:59:03Z","cross_cats_sorted":[],"title_canon_sha256":"c53e7388114353dc17753acb434f7799b4563abced5c0edd7c6d3f6742166908","abstract_canon_sha256":"c69278ea19f1bb98ffe8f9e7c40c0ec650dbcd746b7299449bd7e919e49c780e"},"schema_version":"1.0"},"canonical_sha256":"67d705176d781c1322aaf942668369942bd007edb13abc41477feb2956d4361c","receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T01:05:16.181203Z","signature_b64":"MY6OPInK2EC7SkCRypVestvRBFYYg6mzbFDNesDmRma1w+OEqUZnef+gzFowGUHrAzE59bwcZzQLtlg7xudwAA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"67d705176d781c1322aaf942668369942bd007edb13abc41477feb2956d4361c","last_reissued_at":"2026-05-20T01:05:16.180327Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T01:05:16.180327Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"source_kind":"arxiv","source_id":"2605.08879","source_version":2,"attestation_state":"computed"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-20T01:05:16Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"jDVI3YsrH+i47yq2BhK17pileBCEjsCWklHAo+Uw1KCDom9GYCHtTO9BpmZyDR6qKFkfZ5L0lnVMBylfruhADA==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-27T22:17:24.123753Z"},"content_sha256":"cf4b7c29fb2d3375b6860be88be83d43ccf3983949c4dc50517138c8d91a9316","schema_version":"1.0","event_id":"sha256:cf4b7c29fb2d3375b6860be88be83d43ccf3983949c4dc50517138c8d91a9316"},{"event_type":"graph_snapshot","subject_pith_number":"pith:2026:M7LQKF3NPAOBGIVK7FBGNA3JSQ","target":"graph","payload":{"graph_snapshot":{"paper":{"title":"Preserving Foundational Capabilities in Flow-Matching VLAs through Conservative SFT","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"ConSFT preserves pre-trained capabilities in flow-matching VLAs by scaling learning signals to model confidence during fine-tuning.","cross_cats":[],"primary_cat":"cs.RO","authors_text":"Fuxian Huang, Haoran Zhang, Qi Zhang, Shaopeng Zhai, Tianyi Zhang","submitted_at":"2026-05-09T10:59:03Z","abstract_excerpt":"Unconstrained fine-tuning of flow-matching Vision-Language-Action (VLA) models drives dense parameter overwrites, degrading pre-trained capabilities. We present Conservative Supervised Fine-Tuning (ConSFT), an optimization objective that adapts to target distributions while mitigating catastrophic forgetting, requiring zero prior data or architectural overhead. By dynamically scaling learning signals based on model confidence, ConSFT suppresses excessive gradients from low-confidence samples to prevent disproportionate parameter updates, thereby bounding the intrinsic parameter disruption risk"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"The method outperforms vanilla SFT in capability retention by an average absolute margin of over 20%, matching the efficacy of data-heavy Experience Replay in a prior-data-free regime.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That dynamically scaling learning signals based on model confidence effectively bounds intrinsic parameter disruption risk and secures both target convergence and prior capability retention without introducing new failure modes or limiting necessary adaptation.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"ConSFT prevents catastrophic forgetting in fine-tuning flow-matching VLAs by dynamically scaling gradients based on model confidence, retaining over 20% more pre-trained capability than standard SFT without prior data or reference networks.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"ConSFT preserves pre-trained capabilities in flow-matching VLAs by scaling learning signals to model confidence during fine-tuning.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e9c88f367de35183ce62aa35d394440fbf14139cbabb2eb45d6d2ab001422391"},"source":{"id":"2605.08879","kind":"arxiv","version":2},"verdict":{"id":"a21b10c8-e22f-4561-9088-eec4419d5057","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-12T01:27:22.596915Z","strongest_claim":"The method outperforms vanilla SFT in capability retention by an average absolute margin of over 20%, matching the efficacy of data-heavy Experience Replay in a prior-data-free regime.","one_line_summary":"ConSFT prevents catastrophic forgetting in fine-tuning flow-matching VLAs by dynamically scaling gradients based on model confidence, retaining over 20% more pre-trained capability than standard SFT without prior data or reference networks.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That dynamically scaling learning signals based on model confidence effectively bounds intrinsic parameter disruption risk and secures both target convergence and prior capability retention without introducing new failure modes or limiting necessary adaptation.","pith_extraction_headline":"ConSFT preserves pre-trained capabilities in flow-matching VLAs by scaling learning signals to model confidence during fine-tuning."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.08879/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"ai_meta_artifact","ran_at":"2026-05-19T20:42:28.755374Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T14:01:21.586437Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T10:42:07.085036Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"c09dcb5ca08a22129cee75e5b4efb272cc734a9d6f007108bf430b1c9896df97"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"},"verdict_id":"a21b10c8-e22f-4561-9088-eec4419d5057"},"signer":{"signer_id":"pith.science","signer_type":"pith_registry","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"created_at":"2026-05-20T01:05:16Z","supersedes":[],"prev_event":null,"signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"xZWxHcFes4NEBGdLDh+uzSLoFP5teGC6wONfqRpE0bF3SHbQGTA4ivGe7jOO18idRfPls7Aklx/ETz8tnMgfAQ==","signed_message":"open_graph_event_sha256_bytes","signed_at":"2026-05-27T22:17:24.124730Z"},"content_sha256":"10272a356645849e2c0df19f2b7ce7a430de2006e737e94e9f36a010be979c54","schema_version":"1.0","event_id":"sha256:10272a356645849e2c0df19f2b7ce7a430de2006e737e94e9f36a010be979c54"}],"timestamp_proofs":[],"mirror_hints":[{"mirror_type":"https","name":"Pith Resolver","base_url":"https://pith.science","bundle_url":"https://pith.science/pith/M7LQKF3NPAOBGIVK7FBGNA3JSQ/bundle.json","state_url":"https://pith.science/pith/M7LQKF3NPAOBGIVK7FBGNA3JSQ/state.json","well_known_bundle_url":"https://pith.science/.well-known/pith/M7LQKF3NPAOBGIVK7FBGNA3JSQ/bundle.json","status":"primary"}],"public_keys":[{"key_id":"pith-v1-2026-05","algorithm":"ed25519","format":"raw","public_key_b64":"stVStoiQhXFxp4s2pdzPNoqVNBMojDU/fJ2db5S3CbM=","public_key_hex":"b2d552b68890857171a78b36a5dccf368a953413288c353f7c9d9d6f94b709b3","fingerprint_sha256_b32_first128bits":"RVFV5Z2OI2J3ZUO7ERDEBCYNKS","fingerprint_sha256_hex":"8d4b5ee74e4693bcd1df2446408b0d54","rotates_at":null,"url":"https://pith.science/pith-signing-key.json","notes":"Pith uses this Ed25519 key to sign canonical record SHA-256 digests. Verify with: ed25519_verify(public_key, message=canonical_sha256_bytes, signature=base64decode(signature_b64))."}],"merge_version":"pith-open-graph-merge-v1","built_at":"2026-05-27T22:17:24Z","links":{"resolver":"https://pith.science/pith/M7LQKF3NPAOBGIVK7FBGNA3JSQ","bundle":"https://pith.science/pith/M7LQKF3NPAOBGIVK7FBGNA3JSQ/bundle.json","state":"https://pith.science/pith/M7LQKF3NPAOBGIVK7FBGNA3JSQ/state.json","well_known_bundle":"https://pith.science/.well-known/pith/M7LQKF3NPAOBGIVK7FBGNA3JSQ/bundle.json"},"state":{"state_type":"pith_open_graph_state","state_version":"1.0","pith_number":"pith:2026:M7LQKF3NPAOBGIVK7FBGNA3JSQ","merge_version":"pith-open-graph-merge-v1","event_count":2,"valid_event_count":2,"invalid_event_count":0,"equivocation_count":0,"current":{"canonical_record":{"metadata":{"abstract_canon_sha256":"c69278ea19f1bb98ffe8f9e7c40c0ec650dbcd746b7299449bd7e919e49c780e","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.RO","submitted_at":"2026-05-09T10:59:03Z","title_canon_sha256":"c53e7388114353dc17753acb434f7799b4563abced5c0edd7c6d3f6742166908"},"schema_version":"1.0","source":{"id":"2605.08879","kind":"arxiv","version":2}},"source_aliases":[{"alias_kind":"arxiv","alias_value":"2605.08879","created_at":"2026-05-20T01:05:16Z"},{"alias_kind":"arxiv_version","alias_value":"2605.08879v2","created_at":"2026-05-20T01:05:16Z"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.08879","created_at":"2026-05-20T01:05:16Z"},{"alias_kind":"pith_short_12","alias_value":"M7LQKF3NPAOB","created_at":"2026-05-20T01:05:16Z"},{"alias_kind":"pith_short_16","alias_value":"M7LQKF3NPAOBGIVK","created_at":"2026-05-20T01:05:16Z"},{"alias_kind":"pith_short_8","alias_value":"M7LQKF3N","created_at":"2026-05-20T01:05:16Z"}],"graph_snapshots":[{"event_id":"sha256:10272a356645849e2c0df19f2b7ce7a430de2006e737e94e9f36a010be979c54","target":"graph","created_at":"2026-05-20T01:05:16Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"graph_snapshot":{"author_claims":{"count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","strong_count":0},"builder_version":"pith-number-builder-2026-05-17-v1","claims":{"count":4,"items":[{"attestation":"unclaimed","claim_id":"C1","kind":"strongest_claim","source":"verdict.strongest_claim","status":"machine_extracted","text":"The method outperforms vanilla SFT in capability retention by an average absolute margin of over 20%, matching the efficacy of data-heavy Experience Replay in a prior-data-free regime."},{"attestation":"unclaimed","claim_id":"C2","kind":"weakest_assumption","source":"verdict.weakest_assumption","status":"machine_extracted","text":"That dynamically scaling learning signals based on model confidence effectively bounds intrinsic parameter disruption risk and secures both target convergence and prior capability retention without introducing new failure modes or limiting necessary adaptation."},{"attestation":"unclaimed","claim_id":"C3","kind":"one_line_summary","source":"verdict.one_line_summary","status":"machine_extracted","text":"ConSFT prevents catastrophic forgetting in fine-tuning flow-matching VLAs by dynamically scaling gradients based on model confidence, retaining over 20% more pre-trained capability than standard SFT without prior data or reference networks."},{"attestation":"unclaimed","claim_id":"C4","kind":"headline","source":"verdict.pith_extraction.headline","status":"machine_extracted","text":"ConSFT preserves pre-trained capabilities in flow-matching VLAs by scaling learning signals to model confidence during fine-tuning."}],"snapshot_sha256":"e9c88f367de35183ce62aa35d394440fbf14139cbabb2eb45d6d2ab001422391"},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"integrity":{"available":true,"clean":true,"detectors_run":[{"findings_count":0,"name":"ai_meta_artifact","ran_at":"2026-05-19T20:42:28.755374Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"doi_title_agreement","ran_at":"2026-05-19T14:01:21.586437Z","status":"completed","version":"1.0.0"},{"findings_count":0,"name":"doi_compliance","ran_at":"2026-05-19T10:42:07.085036Z","status":"completed","version":"1.0.0"}],"endpoint":"/pith/2605.08879/integrity.json","findings":[],"snapshot_sha256":"c09dcb5ca08a22129cee75e5b4efb272cc734a9d6f007108bf430b1c9896df97","summary":{"advisory":0,"by_detector":{},"critical":0,"informational":0}},"paper":{"abstract_excerpt":"Unconstrained fine-tuning of flow-matching Vision-Language-Action (VLA) models drives dense parameter overwrites, degrading pre-trained capabilities. We present Conservative Supervised Fine-Tuning (ConSFT), an optimization objective that adapts to target distributions while mitigating catastrophic forgetting, requiring zero prior data or architectural overhead. By dynamically scaling learning signals based on model confidence, ConSFT suppresses excessive gradients from low-confidence samples to prevent disproportionate parameter updates, thereby bounding the intrinsic parameter disruption risk","authors_text":"Fuxian Huang, Haoran Zhang, Qi Zhang, Shaopeng Zhai, Tianyi Zhang","cross_cats":[],"headline":"ConSFT preserves pre-trained capabilities in flow-matching VLAs by scaling learning signals to model confidence during fine-tuning.","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.RO","submitted_at":"2026-05-09T10:59:03Z","title":"Preserving Foundational Capabilities in Flow-Matching VLAs through Conservative SFT"},"references":{"count":0,"internal_anchors":0,"resolved_work":0,"sample":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2605.08879","kind":"arxiv","version":2},"verdict":{"created_at":"2026-05-12T01:27:22.596915Z","id":"a21b10c8-e22f-4561-9088-eec4419d5057","model_set":{"reader":"grok-4.3"},"one_line_summary":"ConSFT prevents catastrophic forgetting in fine-tuning flow-matching VLAs by dynamically scaling gradients based on model confidence, retaining over 20% more pre-trained capability than standard SFT without prior data or reference networks.","pipeline_version":"pith-pipeline@v0.9.0","pith_extraction_headline":"ConSFT preserves pre-trained capabilities in flow-matching VLAs by scaling learning signals to model confidence during fine-tuning.","strongest_claim":"The method outperforms vanilla SFT in capability retention by an average absolute margin of over 20%, matching the efficacy of data-heavy Experience Replay in a prior-data-free regime.","weakest_assumption":"That dynamically scaling learning signals based on model confidence effectively bounds intrinsic parameter disruption risk and secures both target convergence and prior capability retention without introducing new failure modes or limiting necessary adaptation."}},"verdict_id":"a21b10c8-e22f-4561-9088-eec4419d5057"}}],"author_attestations":[],"timestamp_anchors":[],"storage_attestations":[],"citation_signatures":[],"replication_records":[],"corrections":[],"mirror_hints":[],"record_created":{"event_id":"sha256:cf4b7c29fb2d3375b6860be88be83d43ccf3983949c4dc50517138c8d91a9316","target":"record","created_at":"2026-05-20T01:05:16Z","signer":{"key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signer_id":"pith.science","signer_type":"pith_registry"},"payload":{"attestation_state":"computed","canonical_record":{"metadata":{"abstract_canon_sha256":"c69278ea19f1bb98ffe8f9e7c40c0ec650dbcd746b7299449bd7e919e49c780e","cross_cats_sorted":[],"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.RO","submitted_at":"2026-05-09T10:59:03Z","title_canon_sha256":"c53e7388114353dc17753acb434f7799b4563abced5c0edd7c6d3f6742166908"},"schema_version":"1.0","source":{"id":"2605.08879","kind":"arxiv","version":2}},"canonical_sha256":"67d705176d781c1322aaf942668369942bd007edb13abc41477feb2956d4361c","receipt":{"algorithm":"ed25519","builder_version":"pith-number-builder-2026-05-17-v1","canonical_sha256":"67d705176d781c1322aaf942668369942bd007edb13abc41477feb2956d4361c","first_computed_at":"2026-05-20T01:05:16.180327Z","key_id":"pith-v1-2026-05","kind":"pith_receipt","last_reissued_at":"2026-05-20T01:05:16.180327Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","receipt_version":"0.3","signature_b64":"MY6OPInK2EC7SkCRypVestvRBFYYg6mzbFDNesDmRma1w+OEqUZnef+gzFowGUHrAzE59bwcZzQLtlg7xudwAA==","signature_status":"signed_v1","signed_at":"2026-05-20T01:05:16.181203Z","signed_message":"canonical_sha256_bytes"},"source_id":"2605.08879","source_kind":"arxiv","source_version":2}}},"equivocations":[],"invalid_events":[],"applied_event_ids":["sha256:cf4b7c29fb2d3375b6860be88be83d43ccf3983949c4dc50517138c8d91a9316","sha256:10272a356645849e2c0df19f2b7ce7a430de2006e737e94e9f36a010be979c54"],"state_sha256":"ebccbb363ccaa7062a89c1ae7f7c6f727f77ab70829a9a3fd37624d492125625"},"bundle_signature":{"signature_status":"signed_v1","algorithm":"ed25519","key_id":"pith-v1-2026-05","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54","signature_b64":"dIT19bA9V8lwocu1JbeFTX8xLmb6Gj9573SeKgPjEDl7YdX0g2KFnMFdSbKEsHwdym+q0KTlcp/F9AoSQrZzAQ==","signed_message":"bundle_sha256_bytes","signed_at":"2026-05-27T22:17:24.128836Z","bundle_sha256":"e9eb785cf02e4fe215b05c048efb1b78913ab2ec59c89593d227cc33f025e807"}}